SystemError indicates an internal error that is not supposed to be triggerable from Python code. We should probably raise ValueError like plain sockets instead.

>>> s = create_connection(("python.org", 443))
>>> s.recv(-1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: negative buffersize in recv
>>> ss = ssl.wrap_socket(s)
>>> ss.recv(-1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/proj/python/cpython/Lib/ssl.py", line 910, in recv
    return self.read(buflen)
  File "/home/proj/python/cpython/Lib/ssl.py", line 787, in read
    return self._sslobj.read(len, buffer)
  File "/home/proj/python/cpython/Lib/ssl.py", line 573, in read
    v = self._sslobj.read(len or 1024)
SystemError: Negative size passed to PyBytes_FromStringAndSize

Proposed patch

Is this what other file-like objects do with negatives sizes?

Socket objects aren’t exactly file-like. Plain non-SSL sockets don’t even have read() methods.

I think giving a meaning to recv(-1) would be an (unwanted) new feature, rather than a bug fix. If you want a file-like object linked to a socket, I would suggest using something like the makefile() method instead of adding to the low-level socket object API.

But to answer your question: no, most file methods treat a negative size as a special request to read until EOF, e.g. read(-1), readline(-1) and readlines(-1) of RawIOBase, BufferedIOBase and TextIOBase. On the other hand, BufferedIOBase.read1(-1) is poorly defined and supported (Issue 23214), but may end up meaning something like “read an arbitrary non-zero chunk with a minimum amount of low-level calls and processing”.

Thanks for the explanation. Your patch lgtm.

On Sat, Mar 26, 2016, at 15:01, Martin Panter wrote:
> 
> Martin Panter added the comment:
> 
> Socket objects aren’t exactly file-like. Plain non-SSL sockets don’t even
> have read() methods.
> 
> I think giving a meaning to recv(-1) would be an (unwanted) new feature,
> rather than a bug fix. If you want a file-like object linked to a socket,
> I would suggest using something like the makefile() method instead of
> adding to the low-level socket object API.
> 
> But to answer your question: no, most file methods treat a negative size
> as a special request to read until EOF, e.g. read(-1), readline(-1) and
> readlines(-1) of RawIOBase, BufferedIOBase and TextIOBase. On the other
> hand, BufferedIOBase.read1(-1) is poorly defined and supported (Issue
> 23214), but may end up meaning something like “read an arbitrary non-zero
> chunk with a minimum amount of low-level calls and processing”.
> 
> ----------
> 
> _______________________________________
> Python tracker <report@bugs.python.org>
> <http://bugs.python.org/issue26644>
> _______________________________________

Thanks for the reviews. Here is a patch that avoids breaking read(-1, buffer).

LGTM.

New changeset af92651c22e9 by Martin Panter in branch '3.5':
Issue #26644: Raise ValueError for negative SSLSocket.recv() and read()
https://hg.python.org/cpython/rev/af92651c22e9

New changeset b84d136e0028 by Martin Panter in branch '2.7':
Issue #26644: Raise ValueError for negative SSLSocket.recv() and read()
https://hg.python.org/cpython/rev/b84d136e0028

New changeset 80934ad2356d by Martin Panter in branch 'default':
Issue #26644: Merge SSL negative read fix from 3.5
https://hg.python.org/cpython/rev/80934ad2356d

The Python 2 fix was slightly different, due to the lack of Argument Clinic.